What's Changed
- Extend metadata-only bucketing to still images in the discovery backend by @AL3708 in #3001
- implement optimised data filtration for all backends by @bghira in #3165
- support Krea2 turbo LoRA from TheDivergentAI by @bghira in #3168
- Fix Krea2 eval dynamic shift patch sizing by @bghira in #3169
- (#3160) refactor report_to for multiple trackers; remove "all" option by @bghira in #3170
- Add manual system telemetry metrics logging by @bghira in #3172
- Add local metrics galleries and timestep charts by @bghira in #3173
- Prevent training service tests from leaking running jobs by @bghira in #3174
- Fix WebUI save dirty state tracking by @bghira in #3177
- Treat blank publishing config as disabled by @bghira in #3178
- Stabilize cloud hardware profile E2E harness by @bghira in #3179
- Improve training metrics dashboard layout by @bghira in #3180
- Fix validation adapter state restore by @bghira in #3182
- fix(server): five audit_log calls omit the required action argument by @Anai-Guo in #3181
- Improve CUDA OOM log extraction by @bghira in #3183
- Fix adapter-only validation adapter runs by @bghira in #3186
- Fix metrics layout with many validation samples by @bghira in #3187
- Fix validation lightbox checkpoint switching by @bghira in #3189
- Fix W&B tracker initialization handling by @bghira in #3190
- Fix validation artifact asset formats by @bghira in #3191
- fix(quantisation): QuantoLoraConv2d cannot be constructed (update_layer missing use_dora) by @Anai-Guo in #3192
- merge by @bghira in #3193
New Contributors
Full Changelog: v4.9.0...v4.9.1
Contributors
bghira, AL3708, and Anai-Guo